configs/kny_transformer_light.yaml

model:
  transformer_config:
    checkpoint_path: ../../../checkpoints/kny_video_light/checkpoint
    # vocab_size: 50257
    # n_positions: 1024
    # n_embd: 1024 #1280 #768
    # n_layer: 24 #36 #12
    # n_head: 16 #20 #12
    # resid_pdrop: 0.1
    # embd_pdrop: 0.1
    # attn_pdrop: 0.1
    # remaining_frames_method: "concat" 
    # remaining_frames_method: "token_type_ids"
    remaining_frames_method: "own_embeddings"
  first_stage_config:
    checkpoint_path: ../../../checkpoints/kny_image_light_discriminator/checkpoint
    vqvae_config:
      beta: 0.25
      num_embeddings: 64
      embedding_dim: 256
    autoencoder_config:
      z_channels: 128
      channels: 64
      channels_multiplier: 
      - 1
      - 1 
      - 2
      - 2
      - 4
      num_res_blocks: 1
      attention_resolution: 
      - 16
      resolution: 128
      dropout: 0.0
    discriminator_config:
      num_layers: 3
      filters: 64
      
    loss_config:
      discriminator:
        loss: "hinge"
        factor: 1.0
        iter_start: 16200
        weight: 0.3
      vqvae:
        codebook_weight: 1.0
        perceptual_weight: 4.0
      perceptual_loss: "style"

train:
  batch_size: 8
  accumulation_size: 8
  n_epochs: 2000
  len_x_train: 631
  warmup_epoch_percentage: 0.15
  lr_start: 1e-5
  lr_max: 2.5e-4
  perceptual_loss_weight: 1.0
  n_frames_before: 5
  stop_ground_truth_after_epoch: 100